The Discount That Was Real

We have spent a lot of ink on the ways the AI meter runs against you: the cache you can't audit, the tokenizer that quietly inflates the same input, the sticker price that isn't the bill. So it's worth reporting a case that runs the other way, because it's clever, it's real, and it comes with a catch that is very sloppish.

A tool called pxpipe, published to GitHub earlier this month, sits between Claude Code and the API and does something that sounds like a joke: it takes the bulky parts of your context, the system prompt, the tool docs, the older chat history, and renders them into PNG images before sending them, so the model reads them as pictures. Its author reports a 59 to 70% lower end-to-end bill on Fable 5, and about a 60% drop in per-request tokens on the SWE-bench Pro coding benchmark.¹

The reason it works is a seam in how the meter counts. A text token is worth roughly one character. An image token is priced by the picture's pixel dimensions, not by how much text is crammed inside it. On real Claude Code traffic that comes out to about 3.1 characters per image token versus one per text token, so the same passage packed into a picture bills at a fraction of its text cost.¹ Pack your text into a picture and you pay for the frame, not the words.

A text token is worth about one character. An image token is priced by the pixel. That gap is the whole trick.

The catch, which is the story

Here is the part that makes it a receipt rather than a growth hack. Reading text as an image is lossy, and it fails in the worst possible way: silently. The author's own numbers are blunt about it. On a test of recalling a verbatim 12-character hex string from densely rendered content, Opus scored 0 out of 15; Fable 5 managed 13 out of 15. And the misses, in the author's words, are "silent confabulations, not errors."¹ The model does not say it can't read the value. It makes one up and hands it back with full confidence.

So the savings are real, and so is the tax. On the benchmark, the coding score barely moved, 14 of 19 tasks solved with the trick on versus 15 of 19 off,¹ which is why it's tempting. But "barely moved on average" is exactly the shape of a bug that bites you once, on the one request where an API key, a commit hash, or an account number had to be exact, and the model quietly returned a plausible wrong one. You would pay 60% less and never see the invoice for it.

What it actually shows

Take the specific tool out of it and the durable point remains. When you price intelligence by the token, and you count a token differently depending on whether it arrives as text or as pixels, you have created a seam, and someone will always find it and run a truck through it. The pricing model didn't intend for "render your prompt as a JPEG" to be a 60% discount, any more than it intended a new tokenizer to be a stealth price hike. But the unit of account is soft enough that both are true at once.

The meter, it turns out, can be gamed in your favor as easily as it drifts against you. The uncomfortable part is that the cheapest path through it is also the one that most quietly corrupts the thing you were paying for. That's the trade the discount hides: not dollars for nothing, but dollars for the small, silent chance that the answer is confidently wrong.